Kristian Fagerlie describes a retrospective test of eight market-related questions at two points in one week. He says the system used a stored library restricted to information available at each cutoff and that outputs were checked for future leakage. Those safeguards are creator-reported procedures, not independently audited proof of a clean backtest.
Kristian Fagerlie reports that seven of ten simulated positions lost, while a small number of higher-paying outcomes left the experiment ahead overall. The example separates forecasting accuracy from payoff structure and exposes how strongly a short result can depend on one outlier. He explicitly says one week is insufficient and that the system was not designed primarily for such short horizons.
Kristian Fagerlie proposes future tests of linked news events, cross-market effects, continuous monitoring, settlement rules and historical base rates. He presents these as possible sources of forecasting value, not demonstrated advantages. The video does not establish repeatable profitability, transaction-cost-adjusted performance or an investment recommendation.
Kristian Fagerlie demonstrates a custom-question workflow that suggests clearer wording, resolution dates and observable outcome rules before research and parallel forecasts run. Examples include software-industry disruption and an AI-related mathematical announcement. The displayed probabilities and source summaries remain uncertain model outputs; an announcement or favorable expert comment is not itself proof that a mathematical claim is correct.
Watch on YouTube
